[Openais] [corosync] [patch] - Use clock_gettime for timers

[Steven Dake]

Honza,

After more testing, we need the special casing.

MacOSX 10.5.7 (latest version) doesn't support clock_gettime.

Regards
-steve

On Mon, 2009-07-27 at 12:13 +0200, Jan Friesse wrote:
> Committed revision 2373
> 
> Steven Dake napsal(a):
> > Patch looks good for commit
> > 
> > Regards
> > -steve
> > 
> > On Thu, 2009-07-23 at 10:09 +0200, Jan Friesse wrote:
> >> I hope final version of patch.
> >> #ifdefs removed and fixed problem with computing HZ (not used, but may 
> >> be in future)
> >>
> >> Regards,
> >>    Honza
> >>
> >>
> >>> Steve,
> >>> We talked about it on IRC
> >>>
> >>> Jan Friesse wrote:
> >>>> Steve,
> >>>> new patch included.
> >>>>
> >>>> Steven Dake wrote:
> >>>>> In general the patch looks very good.  The issues are as follows:
> >>>>>
> >>>>> 1)
> >>>>>
> >>>>> corosync_timer_add_absolute expects the expire time to be the
> >>>>> nanoseconds since the epoch, or more specifically, 
> >>>>>
> >>>>> from the time(7) man page:
> >>>>>    The Epoch
> >>>>>        Unix  systems  represent  time  in  seconds  since  the Epoch,
> >>>>> which is
> >>>>>        defined as 0:00:00 UTC on the morning of 1 January 1970.
> >>>>>
> >>>>> In this case, the expire_time may not be sufficient.  We want absolute
> >>>>> timers to match the wall clock time rather then be based upon the offset
> >>>>> from the current time.  If that is not the case with your patch, that
> >>>>> needs to be fixed.  absolute timers are used by services to expire
> >>>>> timers when a specific wall clock time has passed.  So in this case, if
> >>>>> ntp moves the time forward, we want that timer to expire, if its
> >>>>> recorded wall clock time is past the current time, the timer should be
> >>>>> expired.
> >>>> This is biggest change. I added is_absolute_timer to timerlist_timer. If
> >>>> this is true -> it's absolute timer and uses old behaviour. If not, it's
> >>>> relative and uses new behaviour. timerlist_add_absolute sets that
> >>>> variable to true, and timerlist_add_duration to false.
> >>>>
> >>>> This is (I hope) how this thing should behave.
> >>>>
> >>>>> 2) souce should be spelled source in configure.ac
> >>>>>
> >>>>> also the logic to detect monotonic clocks in configure.ac seems complex.
> >>>>> I'd like Jim to have a look at the configure.ac to determine if there is
> >>>>> a better way.
> >>>> Typo fixed + Jim added to CC.
> >>>>
> >>>> Jim,
> >>>> I don't think logic is so complex, but maybe there is really some better
> >>>> solution, I'm not autotools expert.
> >>>>
> >>>>> 3) get_monotonic_timeofday used in pload.c and votequorum.c and
> >>>>> totemsrp.c should be in tlist.h and should be named differently.  I'd
> >>>>> choose timerlist_timer_duration_current_get() with same semantics of the
> >>>>> function.
> >>>>>
> >>>>> 4)
> >>>>> an additional function in tlist would be helpful for calculating the
> >>>>> difference in times, such as timerlist_timer_duration_calculate (tv1,
> >>>>> tv2, tv_out)
> >>>> Fixed in way point 7 suggest ;)
> >>>>> 5)
> >>>>> +	nano_monotonic_hz = ts.tv_sec + 1000000000ULL / (unsigned long
> >>>>> long )ts.
> >>>>> experienced c programmers may be able to read this code if they have a
> >>>>> good understanding of operator precedence.  I'd recommend parens to help
> >>>>> clarify to other maintainers the desired order of precedence.
> >>>> Clarified
> >>>>> 6)
> >>>>> +	timer->expire_time= timerlist_nano_monotonic_time() + nano_duration;
> >>>>>
> >>>>> missing space
> >>>> Missing space added
> >>>>
> >>>>> 7)
> >>>>> regarding comment 3 and 4, this could be simplified throughout the code
> >>>>> base by making a new function in timerlist called
> >>>>>
> >>>>> timerlist_nano_current_get() which retrieves the current nanoseconds
> >>>>> from either the epoch or the monotonic clock source depending on os
> >>>>> support.
> >>>>>
> >>>>> this gets rid of all of the timersub junk throughout the code.
> >>>> Done. This makes some bigger changes in computing values, but I hope it
> >>>> makes it clearer. As bonus, I added TIMERLIST_?S_IN_?SEC macros, because
> >>>> writing everywhere values like 1000000000ULL, doesn't look good for me.
> >>>>
> >>>>
> >>>>
> >>>>> 8)
> >>>>> after reviewing the changes, I'm not convinced the absolute timers work
> >>>>> as expected by the coroapi.
> >>>>>
> >>>>> good work
> >>>> I think now will.
> >>>>
> >>>>> Regards
> >>>>> -steve
> >>>>>
> >>>>>
> >>>> Regards,
> >>>>   Honza
> >>>>
> >>>>> On Tue, 2009-07-21 at 15:42 +0200, Jan Friesse wrote:
> >>>>>> Patch should solve problem with ntp and corosync (what it really does)
> >>>>>> by using clock_gettime.
> >>>>>>
> >>>>>> I tried to find EVERY appearance of gettimeofday and where it makes
> >>>>>> sense (so basically, where this value is used as timer) convert to call
> >>>>>> clock_gettime.
> >>>>>>
> >>>>>> Patch modifies configure.ac to detect if clock_gettime and
> >>>>>> CLOCK_MONOTONIC exists on system. If not, gettimeofday will be used (so
> >>>>>> you will get old behavior).
> >>>>>>
> >>>>>> Comments are welcomed.
> >>>>>>
> >>>>>> Regards,
> >>>>>>   Honza
> >>>>>> _______________________________________________
> >>>>>> Openais mailing list
> >>>>>> Openais at lists.linux-foundation.org
> >>>>>> https://lists.linux-foundation.org/mailman/listinfo/openais
> >>>> ------------------------------------------------------------------------
> >>>>
> >>>> _______________________________________________
> >>>> Openais mailing list
> >>>> Openais at lists.linux-foundation.org
> >>>> https://lists.linux-foundation.org/mailman/listinfo/openais
> >>>
> >>> ------------------------------------------------------------------------
> >>>
> >>> _______________________________________________
> >>> Openais mailing list
> >>> Openais at lists.linux-foundation.org
> >>> https://lists.linux-foundation.org/mailman/listinfo/openais
> > 
>